Graphcore and Softbank logos

SoftBank Acquires Beleaguered Graphcore


Graphcore was among the first companies to address data-center AI acceleration and raised more than $700 million. It reported revenue of $10 million in 2019 but less in subsequent years. Sequoia Capital wrote down its stake in 2023, discouraging other investors from pumping in more cash to carry Graphcore through its incubation phase. Willing to place big, risky bets, Softbank has acquired the company, which will continue operating as Graphcore.

Graphcore developed several versions of its AI accelerator (NPU), which it calls an intelligence processing unit (IPU). The company’s architecture instantiates more than 1,000 proprietary RISC cores that have instructions for common activation functions such as ReLu and matrix multiplication and a lot of SRAM. The most recent versions have almost 1 GB on chip. However, the architecture does not support directly attaching off-chip memory but instead requires accessing it via a PCIe-connected gateway IC. Graphcore’s most recent design employed wafer-on-wafer hybrid bonding to attach capacitors to the IPU, improving power delivery and thereby enabling faster clock rates.

Challenging Nvidia at the beginning of the AI wave, Graphcore was present at the right time to become a Top Two data-center AI supplier. The IPU was faster and more power-efficient than Nvidia’s contemporaneous offerings on some workloads but not others. On balance, it wasn’t significantly better, and Graphcore’s software stack wasn’t as complete as Nvidia’s. Thus, the company landed few customers. Meanwhile, its architecture wasn’t well suited to the ever-larger models customers were running, and Nvidia kept scaling up. Demand for an Nvidia alternative remains high, but the investment required to compete is colossal. Beyond its initial purchase price, SoftBank must keep pumping cash into Graphcore to enable it to advance its technology.


Posted

in

by


error: Unable to select